skyloom 1.5.3 → 1.7.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/core/agent.d.ts +2 -0
- package/dist/core/agent.d.ts.map +1 -1
- package/dist/core/agent.js +33 -6
- package/dist/core/agent.js.map +1 -1
- package/dist/core/arbitrate.d.ts +32 -0
- package/dist/core/arbitrate.d.ts.map +1 -0
- package/dist/core/arbitrate.js +136 -0
- package/dist/core/arbitrate.js.map +1 -0
- package/dist/core/estimate.d.ts +30 -0
- package/dist/core/estimate.d.ts.map +1 -0
- package/dist/core/estimate.js +94 -0
- package/dist/core/estimate.js.map +1 -0
- package/dist/core/filter.d.ts +16 -0
- package/dist/core/filter.d.ts.map +1 -0
- package/dist/core/filter.js +91 -0
- package/dist/core/filter.js.map +1 -0
- package/dist/core/index.d.ts +7 -1
- package/dist/core/index.d.ts.map +1 -1
- package/dist/core/index.js +13 -2
- package/dist/core/index.js.map +1 -1
- package/dist/core/learn.d.ts +30 -0
- package/dist/core/learn.d.ts.map +1 -0
- package/dist/core/learn.js +156 -0
- package/dist/core/learn.js.map +1 -0
- package/dist/core/longdoc.d.ts +41 -0
- package/dist/core/longdoc.d.ts.map +1 -0
- package/dist/core/longdoc.js +128 -0
- package/dist/core/longdoc.js.map +1 -0
- package/dist/core/security.d.ts +73 -0
- package/dist/core/security.d.ts.map +1 -0
- package/dist/core/security.js +220 -0
- package/dist/core/security.js.map +1 -0
- package/package.json +1 -1
- package/src/core/agent.ts +18 -6
- package/src/core/arbitrate.ts +162 -0
- package/src/core/estimate.ts +104 -0
- package/src/core/filter.ts +103 -0
- package/src/core/index.ts +8 -2
- package/src/core/learn.ts +146 -0
- package/src/core/longdoc.ts +155 -0
- package/src/core/security.ts +243 -0
|
@@ -0,0 +1,103 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* 输出过滤模块 — sensitive information sanitization.
|
|
3
|
+
*
|
|
4
|
+
* Before agent responses reach the user (or are persisted),
|
|
5
|
+
* scan for and redact sensitive patterns like API keys,
|
|
6
|
+
* tokens, passwords, PII, and internal paths.
|
|
7
|
+
*/
|
|
8
|
+
|
|
9
|
+
/* ═══════════════════════════════════════
|
|
10
|
+
Detection patterns — compiled once at module load
|
|
11
|
+
═══════════════════════════════════════ */
|
|
12
|
+
const SENSITIVE_PATTERNS: Array<[RegExp, string]> = [
|
|
13
|
+
// API keys & tokens
|
|
14
|
+
[/sk-[a-zA-Z0-9]{32,}/g, "[REDACTED:API_KEY]"],
|
|
15
|
+
[/(?:api_key|apikey|secret_key|access_token|auth_token)\s*[:=]\s*["']?[^\s"']{8,}["']?/gi, "$1: [REDACTED]"],
|
|
16
|
+
[/ghp_[a-zA-Z0-9]{36}/g, "[REDACTED:GITHUB_TOKEN]"],
|
|
17
|
+
[/gho_[a-zA-Z0-9]{36}/g, "[REDACTED:GITHUB_TOKEN]"],
|
|
18
|
+
|
|
19
|
+
// AWS credentials
|
|
20
|
+
[/AKIA[0-9A-Z]{16}/g, "[REDACTED:AWS_KEY]"],
|
|
21
|
+
[/(?:aws_access_key_id|aws_secret_access_key)\s*[:=]\s*["']?[^\s"']+/gi, "$1: [REDACTED]"],
|
|
22
|
+
|
|
23
|
+
// Passwords
|
|
24
|
+
[/(?:password|passwd|pwd)\s*[:=]\s*["']?[^\s"']{4,}["']?/gi, "$1: [REDACTED]"],
|
|
25
|
+
[/(?:密码|口令)\s*[:=]\s*["']?[^\s"']{2,}["']?/g, "$1: [已脱敏]"],
|
|
26
|
+
|
|
27
|
+
// Connection strings
|
|
28
|
+
[/(?:mongodb|postgres|mysql|redis):\/\/[^\s]+/g, "[REDACTED:DB_URI]"],
|
|
29
|
+
[/(?:jdbc|odbc):[^\s]+/g, "[REDACTED:DB_URI]"],
|
|
30
|
+
|
|
31
|
+
// Private keys
|
|
32
|
+
[/-----BEGIN (?:RSA |EC |DSA |OPENSSH )?PRIVATE KEY-----[\s\S]*?-----END .*?PRIVATE KEY-----/g, "[REDACTED:PRIVATE_KEY]"],
|
|
33
|
+
|
|
34
|
+
// IP addresses (local only)
|
|
35
|
+
[/192\.168\.\d{1,3}\.\d{1,3}/g, "[REDACTED:LAN_IP]"],
|
|
36
|
+
[/10\.\d{1,3}\.\d{1,3}\.\d{1,3}/g, "[REDACTED:LAN_IP]"],
|
|
37
|
+
[/172\.(1[6-9]|2\d|3[01])\.\d{1,3}\.\d{1,3}/g, "[REDACTED:LAN_IP]"],
|
|
38
|
+
|
|
39
|
+
// File paths
|
|
40
|
+
[/(?:\/etc\/(?:passwd|shadow|hosts|sudoers))/g, "[REDACTED:SYSTEM_PATH]"],
|
|
41
|
+
];
|
|
42
|
+
|
|
43
|
+
/* Email masking (function-based, handled separately) */
|
|
44
|
+
const EMAIL_RE = /([a-zA-Z0-9._%+-]{3,})@([a-zA-Z0-9.-]+\.[a-zA-Z]{2,})/g;
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
/* ═══════════════════════════════════════
|
|
48
|
+
Filter function
|
|
49
|
+
═══════════════════════════════════════ */
|
|
50
|
+
export interface FilterResult {
|
|
51
|
+
clean: string;
|
|
52
|
+
redacted: boolean;
|
|
53
|
+
count: number;
|
|
54
|
+
details: string[];
|
|
55
|
+
}
|
|
56
|
+
|
|
57
|
+
export function filterOutput(text: string): FilterResult {
|
|
58
|
+
if (!text) return { clean: "", redacted: false, count: 0, details: [] };
|
|
59
|
+
|
|
60
|
+
let clean = text;
|
|
61
|
+
let count = 0;
|
|
62
|
+
const details: string[] = [];
|
|
63
|
+
|
|
64
|
+
// Email masking (function-based replacement)
|
|
65
|
+
let emailCount = 0;
|
|
66
|
+
clean = clean.replace(EMAIL_RE, (full, user, domain) => {
|
|
67
|
+
emailCount++;
|
|
68
|
+
return (user as string).slice(0, 2) + "***@" + (domain as string);
|
|
69
|
+
});
|
|
70
|
+
if (emailCount > 0) {
|
|
71
|
+
count += emailCount;
|
|
72
|
+
details.push(`Masked ${emailCount}x email addresses`);
|
|
73
|
+
}
|
|
74
|
+
|
|
75
|
+
for (const [pattern, replacement] of SENSITIVE_PATTERNS) {
|
|
76
|
+
const matches = clean.match(pattern);
|
|
77
|
+
if (matches) {
|
|
78
|
+
count += matches.length;
|
|
79
|
+
if (typeof replacement === "string") {
|
|
80
|
+
details.push(`Redacted ${matches.length}x ${pattern.source.slice(0, 30)}`);
|
|
81
|
+
} else {
|
|
82
|
+
details.push(`Masked ${matches.length}x email addresses`);
|
|
83
|
+
}
|
|
84
|
+
clean = clean.replace(pattern, replacement as string);
|
|
85
|
+
}
|
|
86
|
+
}
|
|
87
|
+
|
|
88
|
+
return { clean, redacted: count > 0, count, details };
|
|
89
|
+
}
|
|
90
|
+
|
|
91
|
+
/* ═══════════════════════════════════════
|
|
92
|
+
Quick check — is filtering needed?
|
|
93
|
+
═══════════════════════════════════════ */
|
|
94
|
+
export function needsFiltering(text: string): boolean {
|
|
95
|
+
if (!text) return false;
|
|
96
|
+
// Quick scan with the most common patterns
|
|
97
|
+
if (/sk-[a-zA-Z0-9]{32,}/.test(text)) return true;
|
|
98
|
+
if (/api_key.*[:=]/.test(text)) return true;
|
|
99
|
+
if (/password.*[:=]/.test(text)) return true;
|
|
100
|
+
if (/-----BEGIN.*PRIVATE KEY-----/.test(text)) return true;
|
|
101
|
+
if (EMAIL_RE.test(text)) return true;
|
|
102
|
+
return false;
|
|
103
|
+
}
|
package/src/core/index.ts
CHANGED
|
@@ -26,6 +26,12 @@ export * from './skill';
|
|
|
26
26
|
export * from './router';
|
|
27
27
|
export * from './agent';
|
|
28
28
|
export * from './factory';
|
|
29
|
+
export * from './security';
|
|
30
|
+
export * from './learn';
|
|
31
|
+
export * from './longdoc';
|
|
32
|
+
export * from './filter';
|
|
33
|
+
export * from './estimate';
|
|
34
|
+
export * from './arbitrate';
|
|
29
35
|
|
|
30
|
-
// Version
|
|
31
|
-
export const VERSION = '1.
|
|
36
|
+
// Version — read from package.json
|
|
37
|
+
export const VERSION = (() => { try { return require('../../package.json').version; } catch { return '1.6.0'; } })();
|
|
@@ -0,0 +1,146 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* 持续学习模块 — post-task review + experience recording.
|
|
3
|
+
*
|
|
4
|
+
* After each task, the agent writes a structured review.
|
|
5
|
+
* Failed attempts are indexed for similarity search to avoid repetition.
|
|
6
|
+
*/
|
|
7
|
+
|
|
8
|
+
import * as fs from "fs";
|
|
9
|
+
import * as path from "path";
|
|
10
|
+
import { USER_CONFIG_DIR } from "./config";
|
|
11
|
+
import { getLogger } from "./logger";
|
|
12
|
+
|
|
13
|
+
const log = getLogger("learn");
|
|
14
|
+
|
|
15
|
+
/* ── Data types ── */
|
|
16
|
+
export interface TaskReview {
|
|
17
|
+
ts: string;
|
|
18
|
+
agent: string;
|
|
19
|
+
goal: string;
|
|
20
|
+
success: boolean;
|
|
21
|
+
durationMs: number;
|
|
22
|
+
toolCalls: string[];
|
|
23
|
+
errorMsg?: string;
|
|
24
|
+
rootCause?: string;
|
|
25
|
+
improvement?: string;
|
|
26
|
+
}
|
|
27
|
+
|
|
28
|
+
export interface ExperienceEntry {
|
|
29
|
+
id: string;
|
|
30
|
+
pattern: string; // What went wrong (key for similarity search)
|
|
31
|
+
solution: string; // What fixed it
|
|
32
|
+
frequency: number; // How often this pattern repeats
|
|
33
|
+
lastSeen: string;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
/* ── Persistence ── */
|
|
37
|
+
const reviewDir = path.join(USER_CONFIG_DIR, "reviews");
|
|
38
|
+
const expFile = path.join(USER_CONFIG_DIR, "experiences.json");
|
|
39
|
+
const reviewDir_ = reviewDir; // for closure
|
|
40
|
+
|
|
41
|
+
function ensureDir() { if (!fs.existsSync(reviewDir_)) fs.mkdirSync(reviewDir_, { recursive: true }); }
|
|
42
|
+
|
|
43
|
+
/* ═══════════════════════════════════════
|
|
44
|
+
Task Review Recording
|
|
45
|
+
═══════════════════════════════════════ */
|
|
46
|
+
export function recordReview(review: TaskReview): void {
|
|
47
|
+
ensureDir();
|
|
48
|
+
const file = path.join(reviewDir_, `${review.ts.slice(0, 10)}_${review.agent}.jsonl`);
|
|
49
|
+
const line = JSON.stringify(review);
|
|
50
|
+
fs.appendFileSync(file, line + "\n");
|
|
51
|
+
log.debug("review_recorded", { agent: review.agent, success: review.success });
|
|
52
|
+
|
|
53
|
+
// If failed, also record as experience
|
|
54
|
+
if (!review.success && review.errorMsg) {
|
|
55
|
+
recordExperience(review.errorMsg, review.rootCause || "unknown", review.improvement || "no improvement noted");
|
|
56
|
+
}
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
/* ═══════════════════════════════════════
|
|
60
|
+
Experience Recording (for failure patterns)
|
|
61
|
+
═══════════════════════════════════════ */
|
|
62
|
+
function loadExperiences(): ExperienceEntry[] {
|
|
63
|
+
try {
|
|
64
|
+
if (fs.existsSync(expFile)) return JSON.parse(fs.readFileSync(expFile, "utf-8"));
|
|
65
|
+
} catch { /* ignore */ }
|
|
66
|
+
return [];
|
|
67
|
+
}
|
|
68
|
+
|
|
69
|
+
function saveExperiences(entries: ExperienceEntry[]): void {
|
|
70
|
+
ensureDir();
|
|
71
|
+
fs.writeFileSync(expFile, JSON.stringify(entries, null, 2), "utf-8");
|
|
72
|
+
}
|
|
73
|
+
|
|
74
|
+
export function recordExperience(errorPattern: string, rootCause: string, solution: string): void {
|
|
75
|
+
const entries = loadExperiences();
|
|
76
|
+
const normalized = errorPattern.toLowerCase().slice(0, 200);
|
|
77
|
+
|
|
78
|
+
// Check for existing similar pattern (simple substring match)
|
|
79
|
+
const existing = entries.find(e => e.pattern.toLowerCase().includes(normalized.slice(0, 50)) || normalized.includes(e.pattern.toLowerCase().slice(0, 50)));
|
|
80
|
+
if (existing) {
|
|
81
|
+
existing.frequency++;
|
|
82
|
+
existing.lastSeen = new Date().toISOString();
|
|
83
|
+
if (solution && solution !== "no improvement noted") existing.solution = solution;
|
|
84
|
+
} else {
|
|
85
|
+
entries.push({
|
|
86
|
+
id: Math.random().toString(36).slice(2, 10),
|
|
87
|
+
pattern: errorPattern.slice(0, 200),
|
|
88
|
+
solution,
|
|
89
|
+
frequency: 1,
|
|
90
|
+
lastSeen: new Date().toISOString(),
|
|
91
|
+
});
|
|
92
|
+
}
|
|
93
|
+
|
|
94
|
+
// Keep top 100 experiences, sorted by frequency
|
|
95
|
+
entries.sort((a, b) => b.frequency - a.frequency);
|
|
96
|
+
if (entries.length > 100) entries.splice(100);
|
|
97
|
+
saveExperiences(entries);
|
|
98
|
+
}
|
|
99
|
+
|
|
100
|
+
/* ═══════════════════════════════════════
|
|
101
|
+
Query experiences
|
|
102
|
+
═══════════════════════════════════════ */
|
|
103
|
+
export function queryExperiences(problem: string, limit: number = 3): ExperienceEntry[] {
|
|
104
|
+
const entries = loadExperiences();
|
|
105
|
+
const lower = problem.toLowerCase();
|
|
106
|
+
return entries
|
|
107
|
+
.filter(e => {
|
|
108
|
+
const plow = e.pattern.toLowerCase();
|
|
109
|
+
// Simple token overlap scoring
|
|
110
|
+
const tokens = lower.split(/\s+/).filter(t => t.length > 2);
|
|
111
|
+
const matches = tokens.filter(t => plow.includes(t));
|
|
112
|
+
return matches.length >= 2;
|
|
113
|
+
})
|
|
114
|
+
.sort((a, b) => b.frequency - a.frequency)
|
|
115
|
+
.slice(0, limit);
|
|
116
|
+
}
|
|
117
|
+
|
|
118
|
+
/* ═══════════════════════════════════════
|
|
119
|
+
Format experiences for system prompt injection
|
|
120
|
+
═══════════════════════════════════════ */
|
|
121
|
+
export function formatExperiencesForPrompt(problem: string): string {
|
|
122
|
+
const exps = queryExperiences(problem);
|
|
123
|
+
if (!exps.length) return "";
|
|
124
|
+
const lines = ["## 历史教训(从经验库检索)", "以下是与当前任务相关的过往失败案例,请避免重复:"];
|
|
125
|
+
for (const e of exps) {
|
|
126
|
+
lines.push(`- **模式**: ${e.pattern.slice(0, 120)}`);
|
|
127
|
+
lines.push(` **解决**: ${e.solution.slice(0, 200)} (出现 ${e.frequency} 次)`);
|
|
128
|
+
}
|
|
129
|
+
return lines.join("\n");
|
|
130
|
+
}
|
|
131
|
+
|
|
132
|
+
/* ═══════════════════════════════════════
|
|
133
|
+
Generate a structured review after task completion
|
|
134
|
+
═══════════════════════════════════════ */
|
|
135
|
+
export function generateReview(
|
|
136
|
+
agent: string, goal: string, success: boolean, durationMs: number,
|
|
137
|
+
toolCalls: string[], errorMsg?: string
|
|
138
|
+
): TaskReview {
|
|
139
|
+
return {
|
|
140
|
+
ts: new Date().toISOString(),
|
|
141
|
+
agent, goal, success, durationMs, toolCalls,
|
|
142
|
+
errorMsg,
|
|
143
|
+
rootCause: errorMsg ? "auto-detected failure" : undefined,
|
|
144
|
+
improvement: errorMsg ? "review error and adjust approach" : undefined,
|
|
145
|
+
};
|
|
146
|
+
}
|
|
@@ -0,0 +1,155 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* 长文档处理策略 — sliding window + summary chain.
|
|
3
|
+
*
|
|
4
|
+
* When an input exceeds the agent's effective context window,
|
|
5
|
+
* split into overlapping chunks, summarize each, then chain
|
|
6
|
+
* summaries into a final digest.
|
|
7
|
+
*
|
|
8
|
+
* Architecture:
|
|
9
|
+
* Input → Chunk(sliding window) → Per-chunk Summary → Chain → Final Digest
|
|
10
|
+
*
|
|
11
|
+
* All summaries are generated by the calling agent's LLM, so quality
|
|
12
|
+
* depends on the model in use. The chunker is pure text processing
|
|
13
|
+
* and works without any LLM call.
|
|
14
|
+
*/
|
|
15
|
+
|
|
16
|
+
import type { BaseAgent } from "./agent";
|
|
17
|
+
|
|
18
|
+
/* ═══════════════════════════════════════
|
|
19
|
+
Chunker — split text into overlapping windows
|
|
20
|
+
═══════════════════════════════════════ */
|
|
21
|
+
export interface ChunkOptions {
|
|
22
|
+
/** Target chunk size in characters (default 6000) */
|
|
23
|
+
chunkSize?: number;
|
|
24
|
+
/** Overlap between consecutive chunks in characters (default 800) */
|
|
25
|
+
overlap?: number;
|
|
26
|
+
/** Minimum chunk size before we stop splitting (default 500) */
|
|
27
|
+
minChunk?: number;
|
|
28
|
+
}
|
|
29
|
+
|
|
30
|
+
export function chunkText(text: string, opts?: ChunkOptions): string[] {
|
|
31
|
+
const cs = opts?.chunkSize ?? 6000;
|
|
32
|
+
const ol = opts?.overlap ?? 800;
|
|
33
|
+
const min = opts?.minChunk ?? 500;
|
|
34
|
+
const chunks: string[] = [];
|
|
35
|
+
|
|
36
|
+
if (text.length <= cs + min) { chunks.push(text); return chunks; }
|
|
37
|
+
|
|
38
|
+
let start = 0;
|
|
39
|
+
while (start < text.length) {
|
|
40
|
+
let end = start + cs;
|
|
41
|
+
if (end >= text.length) { end = text.length; }
|
|
42
|
+
else {
|
|
43
|
+
// Try to break at paragraph boundary
|
|
44
|
+
const searchEnd = Math.min(end + 400, text.length);
|
|
45
|
+
const paraBreak = text.lastIndexOf("\n\n", searchEnd);
|
|
46
|
+
if (paraBreak > start + min) end = paraBreak;
|
|
47
|
+
else {
|
|
48
|
+
const lineBreak = text.lastIndexOf("\n", searchEnd);
|
|
49
|
+
if (lineBreak > start + min) end = lineBreak;
|
|
50
|
+
else {
|
|
51
|
+
const space = text.lastIndexOf(" ", searchEnd);
|
|
52
|
+
if (space > start + min) end = space;
|
|
53
|
+
}
|
|
54
|
+
}
|
|
55
|
+
}
|
|
56
|
+
|
|
57
|
+
chunks.push(text.slice(start, end).trim());
|
|
58
|
+
if (end >= text.length) break;
|
|
59
|
+
start = end - ol;
|
|
60
|
+
if (start < 0) start = 0;
|
|
61
|
+
}
|
|
62
|
+
|
|
63
|
+
return chunks;
|
|
64
|
+
}
|
|
65
|
+
|
|
66
|
+
/* ═══════════════════════════════════════
|
|
67
|
+
Summary chain — ask agent to summarize chunks then chain
|
|
68
|
+
═══════════════════════════════════════ */
|
|
69
|
+
export interface SummaryOptions {
|
|
70
|
+
/** Max total chars for the final digest (default 3000) */
|
|
71
|
+
maxDigestChars?: number;
|
|
72
|
+
/** Custom summarization prompt for each chunk */
|
|
73
|
+
chunkPrompt?: string;
|
|
74
|
+
/** Custom chain prompt for combining summaries */
|
|
75
|
+
chainPrompt?: string;
|
|
76
|
+
}
|
|
77
|
+
|
|
78
|
+
const DEFAULT_CHUNK_PROMPT = `Summarize the following text concisely. Keep all key facts, names, numbers, and code snippets. Output the summary directly without preamble. Limit to 300 words.
|
|
79
|
+
|
|
80
|
+
Text:
|
|
81
|
+
{text}`;
|
|
82
|
+
|
|
83
|
+
const DEFAULT_CHAIN_PROMPT = `Combine the following section summaries into a single coherent digest. Preserve all key facts, remove redundancy. Output directly without preamble.
|
|
84
|
+
|
|
85
|
+
{summaries}`;
|
|
86
|
+
|
|
87
|
+
export async function summarizeLongDoc(
|
|
88
|
+
agent: BaseAgent,
|
|
89
|
+
text: string,
|
|
90
|
+
opts?: SummaryOptions
|
|
91
|
+
): Promise<string> {
|
|
92
|
+
const maxDigest = opts?.maxDigestChars ?? 3000;
|
|
93
|
+
const chunks = chunkText(text);
|
|
94
|
+
|
|
95
|
+
// Single chunk — no summarization needed
|
|
96
|
+
if (chunks.length <= 1) {
|
|
97
|
+
if (text.length <= maxDigest) return text;
|
|
98
|
+
const prompt = (opts?.chunkPrompt || DEFAULT_CHUNK_PROMPT).replace("{text}", text);
|
|
99
|
+
return agent.chatOneshot(prompt, { maxTokens: maxDigest });
|
|
100
|
+
}
|
|
101
|
+
|
|
102
|
+
// Multi-chunk: summarize each, then chain
|
|
103
|
+
const summaries: string[] = [];
|
|
104
|
+
for (let i = 0; i < chunks.length; i++) {
|
|
105
|
+
const prompt = (opts?.chunkPrompt || DEFAULT_CHUNK_PROMPT).replace("{text}", chunks[i]);
|
|
106
|
+
try {
|
|
107
|
+
const s = await agent.chatOneshot(prompt, { maxTokens: 600 });
|
|
108
|
+
summaries.push(s);
|
|
109
|
+
} catch {
|
|
110
|
+
summaries.push(chunks[i].slice(0, 400) + "...");
|
|
111
|
+
}
|
|
112
|
+
}
|
|
113
|
+
|
|
114
|
+
// Chain summaries
|
|
115
|
+
if (summaries.length === 1) return summaries[0].slice(0, maxDigest);
|
|
116
|
+
|
|
117
|
+
const joined = summaries.map((s, i) => `## Section ${i + 1}\n${s}`).join("\n\n");
|
|
118
|
+
if (joined.length <= maxDigest) return joined;
|
|
119
|
+
|
|
120
|
+
const chainPrompt = (opts?.chainPrompt || DEFAULT_CHAIN_PROMPT).replace("{summaries}", joined);
|
|
121
|
+
const final = await agent.chatOneshot(chainPrompt, { maxTokens: maxDigest });
|
|
122
|
+
return final.slice(0, maxDigest);
|
|
123
|
+
}
|
|
124
|
+
|
|
125
|
+
/* ═══════════════════════════════════════
|
|
126
|
+
Structured data parsing helpers
|
|
127
|
+
═══════════════════════════════════════ */
|
|
128
|
+
export function parseStructuredInput(input: string): {
|
|
129
|
+
hasTable: boolean; hasJSON: boolean; hasCSV: boolean;
|
|
130
|
+
extractedJSON: string | null; extractedTable: string[][] | null;
|
|
131
|
+
} {
|
|
132
|
+
const result = { hasTable: false, hasJSON: false, hasCSV: false, extractedJSON: null as string | null, extractedTable: null as string[][] | null };
|
|
133
|
+
|
|
134
|
+
// Detect JSON
|
|
135
|
+
const jsonMatch = input.match(/\{[\s\S]*\}|\[[\s\S]*\]/);
|
|
136
|
+
if (jsonMatch) {
|
|
137
|
+
try { JSON.parse(jsonMatch[0]); result.hasJSON = true; result.extractedJSON = jsonMatch[0]; }
|
|
138
|
+
catch { /* not valid JSON */ }
|
|
139
|
+
}
|
|
140
|
+
|
|
141
|
+
// Detect markdown table
|
|
142
|
+
const tableMatch = input.match(/\|[\s\S]*?\|/);
|
|
143
|
+
if (tableMatch) {
|
|
144
|
+
result.hasTable = true;
|
|
145
|
+
const lines = input.split("\n").filter(l => l.includes("|") && !l.startsWith("|---") && !l.startsWith("| --"));
|
|
146
|
+
result.extractedTable = lines.map(l => l.split("|").filter(c => c.trim()).map(c => c.trim()));
|
|
147
|
+
}
|
|
148
|
+
|
|
149
|
+
// Detect CSV
|
|
150
|
+
if (input.includes(",") && input.split("\n").filter(l => l.includes(",")).length >= 2) {
|
|
151
|
+
result.hasCSV = true;
|
|
152
|
+
}
|
|
153
|
+
|
|
154
|
+
return result;
|
|
155
|
+
}
|
|
@@ -0,0 +1,243 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* 安全与对齐模块 — Security & Alignment
|
|
3
|
+
*
|
|
4
|
+
* Danger level grading, red-line enforcement, audit trail, human-in-the-loop.
|
|
5
|
+
* All security decisions flow through this module before tool execution.
|
|
6
|
+
*/
|
|
7
|
+
|
|
8
|
+
import { getLogger } from "./logger";
|
|
9
|
+
|
|
10
|
+
const log = getLogger("security");
|
|
11
|
+
|
|
12
|
+
/* ── Danger levels ── */
|
|
13
|
+
export enum DangerLevel {
|
|
14
|
+
/** Read-only, no side effects — auto-approved */
|
|
15
|
+
SAFE = 0,
|
|
16
|
+
/** Minor side effects (write single file, git status) — logged */
|
|
17
|
+
LOW = 1,
|
|
18
|
+
/** Significant side effects (overwrite, delete, git push) — notify */
|
|
19
|
+
MEDIUM = 2,
|
|
20
|
+
/** Dangerous (sudo, remote deploy, mass delete) — confirm */
|
|
21
|
+
HIGH = 3,
|
|
22
|
+
/** Red-line — NEVER execute without human-in-the-loop */
|
|
23
|
+
CRITICAL = 4,
|
|
24
|
+
}
|
|
25
|
+
|
|
26
|
+
/* ═══════════════════════════════════════
|
|
27
|
+
Red-line list: operations that are NEVER auto-approved
|
|
28
|
+
═══════════════════════════════════════ */
|
|
29
|
+
const REDLINE_PATTERNS = [
|
|
30
|
+
/rm\s+-rf/, /format\s+\w:/, /dd\s+if=/,
|
|
31
|
+
/>\s*\/dev\/sd/, /mkfs\./, /:(){ :\|:& };:/,
|
|
32
|
+
/sudo\s+rm/, /chmod\s+777\s+\//, /wget.*\|.*sh/,
|
|
33
|
+
/curl.*\|.*bash/, /eval\s+\$/, /exec\s+\$/,
|
|
34
|
+
/subprocess\.call.*rm/, /os\.system.*rm/,
|
|
35
|
+
];
|
|
36
|
+
|
|
37
|
+
const REDLINE_COMMANDS = [
|
|
38
|
+
"shutdown", "reboot", "init 0", "init 6",
|
|
39
|
+
"del /f /s /q C:\\*", "rd /s /q C:\\",
|
|
40
|
+
];
|
|
41
|
+
|
|
42
|
+
/* ═══════════════════════════════════════
|
|
43
|
+
Per-tool danger level mapping
|
|
44
|
+
═══════════════════════════════════════ */
|
|
45
|
+
const TOOL_DANGER_MAP: Record<string, DangerLevel> = {
|
|
46
|
+
read_file: DangerLevel.SAFE,
|
|
47
|
+
list_directory: DangerLevel.SAFE,
|
|
48
|
+
tree: DangerLevel.SAFE,
|
|
49
|
+
file_search: DangerLevel.SAFE,
|
|
50
|
+
code_search: DangerLevel.SAFE,
|
|
51
|
+
grep: DangerLevel.SAFE,
|
|
52
|
+
git_status: DangerLevel.SAFE,
|
|
53
|
+
git_diff: DangerLevel.SAFE,
|
|
54
|
+
git_log: DangerLevel.SAFE,
|
|
55
|
+
system_info: DangerLevel.SAFE,
|
|
56
|
+
system_diagnose: DangerLevel.SAFE,
|
|
57
|
+
list_processes: DangerLevel.SAFE,
|
|
58
|
+
list_installed_apps: DangerLevel.SAFE,
|
|
59
|
+
list_skills: DangerLevel.SAFE,
|
|
60
|
+
recall_facts: DangerLevel.SAFE,
|
|
61
|
+
mcp_list_servers: DangerLevel.SAFE,
|
|
62
|
+
|
|
63
|
+
write_file: DangerLevel.LOW,
|
|
64
|
+
edit_file: DangerLevel.LOW,
|
|
65
|
+
copy_file: DangerLevel.LOW,
|
|
66
|
+
move_file: DangerLevel.LOW,
|
|
67
|
+
http_get: DangerLevel.LOW,
|
|
68
|
+
fetch_page: DangerLevel.LOW,
|
|
69
|
+
web_search: DangerLevel.LOW,
|
|
70
|
+
remember_fact: DangerLevel.LOW,
|
|
71
|
+
use_skill: DangerLevel.LOW,
|
|
72
|
+
task_done: DangerLevel.LOW,
|
|
73
|
+
|
|
74
|
+
delete_file: DangerLevel.MEDIUM,
|
|
75
|
+
git_add: DangerLevel.MEDIUM,
|
|
76
|
+
git_commit: DangerLevel.MEDIUM,
|
|
77
|
+
git_checkout: DangerLevel.MEDIUM,
|
|
78
|
+
http_post: DangerLevel.MEDIUM,
|
|
79
|
+
mcp_add_server: DangerLevel.MEDIUM,
|
|
80
|
+
mcp_remove_server: DangerLevel.MEDIUM,
|
|
81
|
+
launch_app: DangerLevel.MEDIUM,
|
|
82
|
+
open_path: DangerLevel.MEDIUM,
|
|
83
|
+
browser_open: DangerLevel.MEDIUM,
|
|
84
|
+
|
|
85
|
+
run_bash: DangerLevel.HIGH,
|
|
86
|
+
shell_exec: DangerLevel.HIGH,
|
|
87
|
+
kill_process: DangerLevel.HIGH,
|
|
88
|
+
package_manager: DangerLevel.HIGH,
|
|
89
|
+
service_control: DangerLevel.HIGH,
|
|
90
|
+
delegate_to: DangerLevel.HIGH,
|
|
91
|
+
mcp_scaffold_server: DangerLevel.HIGH,
|
|
92
|
+
};
|
|
93
|
+
|
|
94
|
+
/* ═══════════════════════════════════════
|
|
95
|
+
Audit trail entry
|
|
96
|
+
═══════════════════════════════════════ */
|
|
97
|
+
export interface AuditEntry {
|
|
98
|
+
ts: string;
|
|
99
|
+
agent: string;
|
|
100
|
+
tool: string;
|
|
101
|
+
args: Record<string, any>;
|
|
102
|
+
dangerLevel: DangerLevel;
|
|
103
|
+
approved: boolean;
|
|
104
|
+
result: string;
|
|
105
|
+
durationMs: number;
|
|
106
|
+
traceId: string;
|
|
107
|
+
}
|
|
108
|
+
|
|
109
|
+
/* ═══════════════════════════════════════
|
|
110
|
+
Security context — per-session security state
|
|
111
|
+
═══════════════════════════════════════ */
|
|
112
|
+
export class SecurityContext {
|
|
113
|
+
public auditLog: AuditEntry[] = [];
|
|
114
|
+
public deniedCount = 0;
|
|
115
|
+
public autoApprovedCount = 0;
|
|
116
|
+
public manualApprovedCount = 0;
|
|
117
|
+
public approvalMode: "auto" | "interactive" | "strict" = "auto";
|
|
118
|
+
|
|
119
|
+
private approvalCallback: ((tool: string, args: Record<string, any>, level: DangerLevel) => Promise<boolean>) | null = null;
|
|
120
|
+
|
|
121
|
+
constructor(opts?: { mode?: "auto" | "interactive" | "strict"; onApprove?: (tool: string, args: Record<string, any>, level: DangerLevel) => Promise<boolean> }) {
|
|
122
|
+
if (opts?.mode) this.approvalMode = opts.mode;
|
|
123
|
+
if (opts?.onApprove) this.approvalCallback = opts.onApprove;
|
|
124
|
+
}
|
|
125
|
+
|
|
126
|
+
/** Get the danger level for a tool. Defaults to SAFE for unknown tools. */
|
|
127
|
+
getDangerLevel(toolName: string): DangerLevel {
|
|
128
|
+
return TOOL_DANGER_MAP[toolName] ?? DangerLevel.SAFE;
|
|
129
|
+
}
|
|
130
|
+
|
|
131
|
+
/** Check if arguments contain red-line patterns (critical danger). */
|
|
132
|
+
checkRedline(toolName: string, args: Record<string, any>): string | null {
|
|
133
|
+
if (toolName !== "run_bash" && toolName !== "shell_exec") return null;
|
|
134
|
+
const cmd = String(args.command || args.cmd || "").toLowerCase();
|
|
135
|
+
for (const pattern of REDLINE_PATTERNS) {
|
|
136
|
+
if (pattern.test(cmd)) return `Red-line pattern detected: ${pattern.source.slice(0, 40)}`;
|
|
137
|
+
}
|
|
138
|
+
for (const forbidden of REDLINE_COMMANDS) {
|
|
139
|
+
if (cmd.includes(forbidden)) return `Red-line command: ${forbidden}`;
|
|
140
|
+
}
|
|
141
|
+
return null;
|
|
142
|
+
}
|
|
143
|
+
|
|
144
|
+
/** Determine whether a tool call is permitted. Returns [approved, reason]. */
|
|
145
|
+
async checkApproval(toolName: string, args: Record<string, any>, agentName: string): Promise<[boolean, string]> {
|
|
146
|
+
const level = this.getDangerLevel(toolName);
|
|
147
|
+
|
|
148
|
+
// Red-line check
|
|
149
|
+
const redline = this.checkRedline(toolName, args);
|
|
150
|
+
if (redline) {
|
|
151
|
+
log.warn("redline_blocked", { agent: agentName, tool: toolName, reason: redline });
|
|
152
|
+
return [false, redline];
|
|
153
|
+
}
|
|
154
|
+
|
|
155
|
+
// Safe — always allow
|
|
156
|
+
if (level === DangerLevel.SAFE) return [true, "safe"];
|
|
157
|
+
|
|
158
|
+
// Strict mode — deny all non-safe
|
|
159
|
+
if (this.approvalMode === "strict") {
|
|
160
|
+
return [false, `Strict mode: tool '${toolName}' (level ${level}) requires manual approval`];
|
|
161
|
+
}
|
|
162
|
+
|
|
163
|
+
// Auto mode — allow LOW, prompt for MEDIUM+, deny CRITICAL
|
|
164
|
+
if (this.approvalMode === "auto") {
|
|
165
|
+
if (level <= DangerLevel.LOW) return [true, "auto-low"];
|
|
166
|
+
if (level === DangerLevel.CRITICAL) return [false, `CRITICAL tool '${toolName}' requires explicit human approval`];
|
|
167
|
+
// MEDIUM/HIGH with auto mode => need callback
|
|
168
|
+
if (this.approvalCallback) {
|
|
169
|
+
const approved = await this.approvalCallback(toolName, args, level);
|
|
170
|
+
return [approved, approved ? "user-approved" : "user-denied"];
|
|
171
|
+
}
|
|
172
|
+
return [true, "auto-med"]; // no callback → auto-allow but log
|
|
173
|
+
}
|
|
174
|
+
|
|
175
|
+
// Interactive mode — prompt for LOW+
|
|
176
|
+
if (this.approvalCallback) {
|
|
177
|
+
const approved = await this.approvalCallback(toolName, args, level);
|
|
178
|
+
return [approved, approved ? "user-approved" : "user-denied"];
|
|
179
|
+
}
|
|
180
|
+
return [true, "no-callback"];
|
|
181
|
+
}
|
|
182
|
+
|
|
183
|
+
/** Record an audit entry. */
|
|
184
|
+
recordAudit(tool: string, agent: string, args: Record<string, any>, dangerLevel: DangerLevel, approved: boolean, resultPreview: string, durationMs: number, traceId: string): void {
|
|
185
|
+
const entry: AuditEntry = {
|
|
186
|
+
ts: new Date().toISOString(),
|
|
187
|
+
agent, tool, args, dangerLevel, approved,
|
|
188
|
+
result: resultPreview.slice(0, 500),
|
|
189
|
+
durationMs, traceId,
|
|
190
|
+
};
|
|
191
|
+
this.auditLog.push(entry);
|
|
192
|
+
if (this.auditLog.length > 5000) this.auditLog.shift();
|
|
193
|
+
|
|
194
|
+
if (approved) {
|
|
195
|
+
if (dangerLevel >= DangerLevel.HIGH) this.manualApprovedCount++;
|
|
196
|
+
else this.autoApprovedCount++;
|
|
197
|
+
} else {
|
|
198
|
+
this.deniedCount++;
|
|
199
|
+
}
|
|
200
|
+
|
|
201
|
+
log.info(dangerLevel >= DangerLevel.HIGH ? "dangerous_tool_executed" : "tool_executed", {
|
|
202
|
+
tool, agent, level: dangerLevel, approved,
|
|
203
|
+
});
|
|
204
|
+
}
|
|
205
|
+
|
|
206
|
+
/** Get summary statistics. */
|
|
207
|
+
getStats() {
|
|
208
|
+
return {
|
|
209
|
+
total: this.auditLog.length,
|
|
210
|
+
denied: this.deniedCount,
|
|
211
|
+
autoApproved: this.autoApprovedCount,
|
|
212
|
+
manualApproved: this.manualApprovedCount,
|
|
213
|
+
byLevel: {
|
|
214
|
+
safe: this.auditLog.filter(e => e.dangerLevel === DangerLevel.SAFE).length,
|
|
215
|
+
low: this.auditLog.filter(e => e.dangerLevel === DangerLevel.LOW).length,
|
|
216
|
+
medium: this.auditLog.filter(e => e.dangerLevel === DangerLevel.MEDIUM).length,
|
|
217
|
+
high: this.auditLog.filter(e => e.dangerLevel === DangerLevel.HIGH).length,
|
|
218
|
+
critical: this.auditLog.filter(e => e.dangerLevel === DangerLevel.CRITICAL).length,
|
|
219
|
+
},
|
|
220
|
+
lastDenied: this.auditLog.filter(e => !e.approved).slice(-5).map(e => `${e.tool}: ${e.result}`),
|
|
221
|
+
};
|
|
222
|
+
}
|
|
223
|
+
|
|
224
|
+
/** Install approval callback for interactive mode. */
|
|
225
|
+
setApprovalCallback(fn: (tool: string, args: Record<string, any>, level: DangerLevel) => Promise<boolean>) {
|
|
226
|
+
this.approvalCallback = fn;
|
|
227
|
+
}
|
|
228
|
+
}
|
|
229
|
+
|
|
230
|
+
/* ── Global security context ── */
|
|
231
|
+
let globalSecurity: SecurityContext | null = null;
|
|
232
|
+
|
|
233
|
+
export function getSecurity(): SecurityContext {
|
|
234
|
+
if (!globalSecurity) globalSecurity = new SecurityContext();
|
|
235
|
+
return globalSecurity;
|
|
236
|
+
}
|
|
237
|
+
|
|
238
|
+
export function resetSecurity(): void {
|
|
239
|
+
globalSecurity = null;
|
|
240
|
+
}
|
|
241
|
+
|
|
242
|
+
/** Red-line patterns for reference (used by tools to self-check). */
|
|
243
|
+
export { REDLINE_PATTERNS, REDLINE_COMMANDS };
|